如何将 fit_generator 与多个输入一起使用

Question

是否可以有两个fit_generator？

我正在创建一个有两个输入的模型， 模型配置如下图。

标签 Y 对 X1 和 X2 数据使用相同的标签。

以下错误会继续出现。

检查模型输入时出错：您传递给模型的 Numpy 数组列表不是模型预期的大小。预期的 查看 2 个数组，但得到了以下 1 个数组的列表： [数组([[[[0.75686276, 0.75686276, 0.75686276], [0.75686276, 0.75686276, 0.75686276], [0.75686276, 0.75686276, 0.75686276], ..., [0.65882355, 0.65882355, 0.65882355...

我的代码如下所示：

def generator_two_img(X1, X2, Y,batch_size):
    generator = ImageDataGenerator(rotation_range=15,
                                   width_shift_range=0.2,
                                   height_shift_range=0.2,
                                   shear_range=0.2,
                                   zoom_range=0.2,
                                   horizontal_flip=True,
                                   fill_mode='nearest')

    genX1 = generator.flow(X1, Y, batch_size=batch_size)
    genX2 = generator.flow(X2, Y, batch_size=batch_size)

    while True:
        X1 = genX1.__next__()
        X2 = genX2.__next__()
        yield [X1, X2], Y
  """
      .................................
  """
hist = model.fit_generator(generator_two_img(x_train, x_train_landmark, 
                y_train, batch_size),
                steps_per_epoch=len(x_train) // batch_size, epochs=nb_epoch,
                callbacks = callbacks,
                validation_data=(x_validation, y_validation),
                validation_steps=x_validation.shape[0] // batch_size, 
                `enter code here`verbose=1)

Answer 1

试试这个生成器：

def generator_two_img(X1, X2, y, batch_size):
    genX1 = gen.flow(X1, y,  batch_size=batch_size, seed=1)
    genX2 = gen.flow(X2, y, batch_size=batch_size, seed=1)
    while True:
        X1i = genX1.next()
        X2i = genX2.next()
        yield [X1i[0], X2i[0]], X1i[1]

3 个输入的生成器：

def generator_three_img(X1, X2, X3, y, batch_size):
    genX1 = gen.flow(X1, y,  batch_size=batch_size, seed=1)
    genX2 = gen.flow(X2, y, batch_size=batch_size, seed=1)
    genX3 = gen.flow(X3, y, batch_size=batch_size, seed=1)
    while True:
        X1i = genX1.next()
        X2i = genX2.next()
        X3i = genX3.next()
        yield [X1i[0], X2i[0], X3i[0]], X1i[1]

Answer 2

我有一个针对TimeseriesGenerator 的多个输入的实现，我已经对其进行了调整（不幸的是，我无法对其进行测试）以使用ImageDataGenerator 来满足这个示例。我的方法是从keras.utils.Sequence 为多个生成器构建一个包装类，然后实现它的基本方法：__len__ 和__getitem__：

from keras.preprocessing.image import ImageDataGenerator
from keras.utils import Sequence


class MultipleInputGenerator(Sequence):
    """Wrapper of 2 ImageDataGenerator"""

    def __init__(self, X1, X2, Y, batch_size):
        # Keras generator
        self.generator = ImageDataGenerator(rotation_range=15, 
                                            width_shift_range=0.2,
                                            height_shift_range=0.2,
                                            shear_range=0.2,
                                            zoom_range=0.2,
                                            horizontal_flip=True, 
                                            fill_mode='nearest')

        # Real time multiple input data augmentation
        self.genX1 = self.generator.flow(X1, Y, batch_size=batch_size)
        self.genX2 = self.generator.flow(X2, Y, batch_size=batch_size)

    def __len__(self):
        """It is mandatory to implement it on Keras Sequence"""
        return self.genX1.__len__()

    def __getitem__(self, index):
        """Getting items from the 2 generators and packing them"""
        X1_batch, Y_batch = self.genX1.__getitem__(index)
        X2_batch, Y_batch = self.genX2.__getitem__(index)

        X_batch = [X1_batch, X2_batch]

        return X_batch, Y_batch

生成器实例化后，您可以将此生成器与model.fit_generator() 一起使用。